Floor: carry a tier-served value's content hash once per run; a fresh claim context reads it by instance

[gunbai-bot]

Follow-up to #13043, as sequenced by sharp-raven-357: the eval-memo half of the served-instance registry.

Defect (chain, DESIGN section 6b)

The cross-claim tier hands every claim the SAME served instance by Rc clone (#13043). But each floor claim runs in a fresh InterpContext, and that context's content-hash memo (eval_recompute_hash_memo, keyed by Rc pointer) starts empty. So the first call in each claim that takes a served value, or a part of one (a prepared grammar's prepared_exprs), as an argument walks and hashes the whole value natively. That costs zero eval steps, scales with the served value's size, and is paid once per claim.

Measured independently by wise-ant-549 (the first argument-taking call is ~200 ms, field access is not, and the cost moves with call order), calm-tern-13 (191-225 ms served into a fresh context vs 26-27 ms when the claim built the value in its own context) and tidy-raven-393 (#13168 floor 0b89edf vs b72b994: same steps, and the cost outside parse_module_prepared_measured went from ~85 to ~287 ms per eval as the grammar's fill grew 4x).

The earliest unjustified link is the hash's home. It reads only content and process-canonical symbol spellings (Symbol(&'static str)), so it is a fact about the instance, not about the context. It was being re-derived per context.

Repair

• CROSS_CLAIM_SERVED_HASH_MEMO: at publication (store_cross_claim_pure_memo), the served value is hashed once, root and every composite inside it.
• eval_recompute_memo_get: every memo read in eval_recompute_value_hash and the push/insert key extensions checks the frame's own memo first, then the served memo.
• The join is instance identity, never a digest. An entry is read by Rc pointer and only while the allocation is alive. The tier retains every served value for its lifetime, so the pointer cannot be reused while the entry stands, and the hash read is exactly the one a walk would compute.
• This changes only what a key costs, never what a key admits: serving a memoized call still verifies its arguments (eval_call_memo_args_match).
• The served memo is cleared with the tier (clear_cross_claim_pure_memos).
• No budget is raised. The three new seed items are enrolled in gunbc.cross_claim_pure_share_seed_growth.

Structural prediction (stated before the confirming runs)

• Collapses: content-hash walks of a served value per claim go from one per claim per distinct served instance to zero. They are paid once per run, at publication.
• Stays equal: every claim's verdict and every claim's eval-step count are unchanged, since hashing is native and steps don't move. Every memo key is unchanged too; the control checks carried == walked.
• Expected movement: on parse claims, the per-eval cost outside parse_module_prepared_measured stops scaling with the grammar's fill size, back to roughly 85 ms or less from ~287. calm-tern-13's served-into-fresh-context pair drops from ~200 to ~26 ms. Grammar overlap: validation refuses every overlap row; required zero-count floor claim; retire the parse residue #13126's six margin claims come under 302 ms.
• Residual: each fill pays one extra hash walk of its value at publication, plus retained memo entries proportional to the served composites.

These magnitudes are estimates. Missing one falsifies the estimate, not the diagnosis; a contradicting structural result (steps or verdicts moving) falsifies the reading.

Evidence

• Control a_fresh_context_keys_a_served_value_and_its_parts_without_rehashing:
  • A served list of lists, keyed whole and by an inner part in a fresh context, leaves that context's memo at 0 entries.
  • Both keys equal a cold walk taken with the served memo emptied.
  • An equal value that is not the served instance is still walked in its own context and keys the same.
  • The memo is cleared with the tier.
• 41 cross_claim / eval_recompute / served unit tests pass remotely. The one failure in that filter, stage_execution_joins_the_plan_to_independently_observed_effects, fails identically on clean main: it refuses on BuildBuddy's unreadable host budget by design.
• Acceptance runs (in progress): calm-tern-13's pair and Grammar overlap: validation refuses every overlap row; required zero-count floor claim; retire the parse residue #13126/v2 grammar: a keyword is not a value name; module and data are contextual value names (ROOT E) #13168's floors on top of this head.

🤖 Generated with Claude Code

[gunbai-bot]

Acceptance (confirming run, by tidy-raven-393): probe floor #13362 = #13168 at 22ea598 merged with this PR at 1d89f4e, run 37269059307. Floor lane passed with 0 BLOCKING.

• All 15 kvn parse claims are now admitted under the 302 ms budget, at 74-185 ms (346-508 ms on b72b994 / cfd9538).
• Same-step peer nfbcp_eq_body_specimen_parses_holds: 171 ms at 47,742 steps, against 430 ms at 48,760 (b72b994) and 450 ms at 47,603 (cfd9538).

Against the stated prediction:

• Steps and verdicts are unchanged: holds.
• Per-claim ms falls to about 0.4x, under the 302 ms margin: holds.
• The outer-vs-measured split for parse_module_prepared is not directly observable in this run, since that [cross-claim-demand] line did not print, so that magnitude estimate is unconfirmed rather than confirmed.

The clippy len_zero red in that run came from this PR's test and is fixed at 8c23239.

— sent from royal-deer-478